Materials Process Design and Control Laboratory 1 An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation.

Slides:



Advertisements
Similar presentations
Pattern Recognition and Machine Learning
Advertisements

Introduction to Monte Carlo Markov chain (MCMC) methods
Bayesian Belief Propagation
Point-wise Discretization Errors in Boundary Element Method for Elasticity Problem Bart F. Zalewski Case Western Reserve University Robert L. Mullen Case.
Bayesian Estimation in MARK
1 12. Principles of Parameter Estimation The purpose of this lecture is to illustrate the usefulness of the various concepts introduced and studied in.
EARS1160 – Numerical Methods notes by G. Houseman
CHAPTER 16 MARKOV CHAIN MONTE CARLO
Bayesian statistics – MCMC techniques
Visual Recognition Tutorial
Pattern Recognition and Machine Learning
2-1 Sample Spaces and Events Conducting an experiment, in day-to-day repetitions of the measurement the results can differ slightly because of small.
Machine Learning CUNY Graduate Center Lecture 3: Linear Regression.
Using ranking and DCE data to value health states on the QALY scale using conventional and Bayesian methods Theresa Cain.
Copyright ©2011 Pearson Education 8-1 Chapter 8 Confidence Interval Estimation Statistics for Managers using Microsoft Excel 6 th Global Edition.
Computer vision: models, learning and inference Chapter 10 Graphical Models.
Lecture II-2: Probability Review
Robin McDougall, Ed Waller and Scott Nokleby Faculties of Engineering & Applied Science and Energy Systems & Nuclear Science 1.
CSC2535: 2013 Advanced Machine Learning Lecture 3a: The Origin of Variational Bayes Geoffrey Hinton.
Radial Basis Function Networks
Advanced Image Processing Image Relaxation – Restoration and Feature Extraction 02/02/10.
Image Analysis and Markov Random Fields (MRFs) Quanren Xiong.
Reduced-order modeling of stochastic transport processes Materials Process Design and Control Laboratory Swagato Acharjee and Nicholas Zabaras Materials.
Binary Variables (1) Coin flipping: heads=1, tails=0 Bernoulli Distribution.
Physics 114: Lecture 15 Probability Tests & Linear Fitting Dale E. Gary NJIT Physics Department.
Component Reliability Analysis
Ch 8.1 Numerical Methods: The Euler or Tangent Line Method
Gaussian process modelling
An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method Xiang Ma and Nicholas Zabaras Materials.
1 Hybrid methods for solving large-scale parameter estimation problems Carlos A. Quintero 1 Miguel Argáez 1 Hector Klie 2 Leticia Velázquez 1 Mary Wheeler.
Materials Process Design and Control Laboratory THE STEFAN PROBLEM: A STOCHASTIC ANALYSIS USING THE EXTENDED FINITE ELEMENT METHOD Baskar Ganapathysubramanian,
Statistical Decision Theory
Statistics for Managers Using Microsoft Excel, 5e © 2008 Pearson Prentice-Hall, Inc.Chap 8-1 Statistics for Managers Using Microsoft® Excel 5th Edition.
Module 1: Statistical Issues in Micro simulation Paul Sousa.
ECE 8443 – Pattern Recognition ECE 8423 – Adaptive Signal Processing Objectives: Deterministic vs. Random Maximum A Posteriori Maximum Likelihood Minimum.
Finding Scientific topics August , Topic Modeling 1.A document as a probabilistic mixture of topics. 2.A topic as a probability distribution.
Materials Process Design and Control Laboratory 1 High-dimensional model representation technique for the solution of stochastic PDEs Nicholas Zabaras.
ECE 8443 – Pattern Recognition LECTURE 07: MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION Objectives: Class-Conditional Density The Multivariate Case General.
Materials Process Design and Control Laboratory 1 Stochastic Modeling in High-Dimensional Spaces Nicholas Zabaras Materials Process Design and Control.
Fast Simulators for Assessment and Propagation of Model Uncertainty* Jim Berger, M.J. Bayarri, German Molina June 20, 2001 SAMO 2001, Madrid *Project of.
Uncertainty quantification in multiscale deformation processes Babak Kouchmeshky Nicholas Zabaras Materials Process Design and Control Laboratory Sibley.
PROBABILITY AND STATISTICS FOR ENGINEERING Hossein Sameti Department of Computer Engineering Sharif University of Technology Principles of Parameter Estimation.
Materials Process Design and Control Laboratory 1 An adaptive hierarchical sparse grid collocation algorithm for the solution of stochastic differential.
Statistical Decision Theory Bayes’ theorem: For discrete events For probability density functions.
ECE-7000: Nonlinear Dynamical Systems Overfitting and model costs Overfitting  The more free parameters a model has, the better it can be adapted.
HEAT TRANSFER FINITE ELEMENT FORMULATION
CHAPTER 17 O PTIMAL D ESIGN FOR E XPERIMENTAL I NPUTS Organization of chapter in ISSO –Background Motivation Finite sample and asymptotic (continuous)
Stats 845 Applied Statistics. This Course will cover: 1.Regression –Non Linear Regression –Multiple Regression 2.Analysis of Variance and Experimental.
1  The Problem: Consider a two class task with ω 1, ω 2   LINEAR CLASSIFIERS.
- 1 - Overall procedure of validation Calibration Validation Figure 12.4 Validation, calibration, and prediction (Oberkampf and Barone, 2004 ). Model accuracy.
Sampling and estimation Petter Mostad
Chapter 20 Classification and Estimation Classification – Feature selection Good feature have four characteristics: –Discrimination. Features.
6. Population Codes Presented by Rhee, Je-Keun © 2008, SNU Biointelligence Lab,
Lecture 3: MLE, Bayes Learning, and Maximum Entropy
1 Chapter 8: Model Inference and Averaging Presented by Hui Fang.
Lecture II-3: Interpolation and Variational Methods Lecture Outline: The Interpolation Problem, Estimation Options Regression Methods –Linear –Nonlinear.
Materials Process Design and Control Laboratory Sibley School of Mechanical and Aerospace Engineering 169 Frank H. T. Rhodes Hall Cornell University Ithaca,
Materials Process Design and Control Laboratory 1 A stochastic dimension reduction for stochastic PDEs Nicholas Zabaras Materials Process Design and Control.
Kevin Stevenson AST 4762/5765. What is MCMC?  Random sampling algorithm  Estimates model parameters and their uncertainty  Only samples regions of.
Ch 1. Introduction Pattern Recognition and Machine Learning, C. M. Bishop, Updated by J.-H. Eom (2 nd round revision) Summarized by K.-I.
Fundamentals of Data Analysis Lecture 11 Methods of parametric estimation.
Biointelligence Laboratory, Seoul National University
T. E. Dyhoum1, D. Lesnic 1 and R. G. Aykroyd 2
CSCI 5822 Probabilistic Models of Human and Machine Learning
Remember that our objective is for some density f(y|) for observations where y and  are vectors of data and parameters,  being sampled from a prior.
Filtering and State Estimation: Basic Concepts
Where did we stop? The Bayes decision rule guarantees an optimal classification… … But it requires the knowledge of P(ci|x) (or p(x|ci) and P(ci)) We.
Parametric Methods Berlin Chen, 2005 References:
Yalchin Efendiev Texas A&M University
Probabilistic Surrogate Models
Presentation transcript:

Materials Process Design and Control Laboratory 1 An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method Nicholas Zabaras and Xiang Ma Materials Process Design and Control Laboratory Sibley School of Mechanical and Aerospace Engineering 101 Frank H. T. Rhodes Hall Cornell University Ithaca, NY URL: CSE09, SIAM Conference on Computational Science and Engineering, Miami, FL, March 2-6, 2009

Materials Process Design and Control Laboratory 2 Outline of the presentation  Bayesian approach to inverse problems  Accelerating Bayesian inference via adaptive sparse grid collocation method  Inference of spatial field ---- MRF, Process convolution model, Regularization  Conclusions

Materials Process Design and Control Laboratory 3 Inverse problems system input (unknowns) inverse problem: effect(d)–to-cause(m) mathematical representation d = G(m, ω s ) + ω m physical processes governed by PDEs observations forward problem: cause(m)-to-effect(d) MELT SOLID G,V g Forward operators e.g. a system of PDEs; computationally intensive observed data model parameters/inputs, assume random Given a set of data d, estimate m

Materials Process Design and Control Laboratory 4 Applications and significance Inverse heat transfer permeability estimation of a reservoir ODF pole figure microstructure materials process design and control groundwater cleanup broad applications

Materials Process Design and Control Laboratory 5 Features of inverse problem ill-posedness well-posedness: --- existence --- uniqueness --- continuous dependence of solutions on measurements stability inverse problem is ill-posed identifiability implicit objective function (functional) non-linearity complex direct simulation high computation cost uncertainties

Materials Process Design and Control Laboratory 6 Bayesian computational approach Obstacle: Posteriors are often high-dimensional Computation cost for repeated likelihood evaluations aim: explore the advantages and resolve this difficulty provides probabilistic description of the inverse solution direct simulation in deterministic space prior distribution regularization --- incorporation of prior knowledge --- explores the distribution of regularization parameter Advantages:

Materials Process Design and Control Laboratory 7 Bayesian inference  The model m is now a random variable/field.  Apply Bayes’ theorem: posterior density Likelihood function L (m) prior density Normalizing const The posterior density is the full Bayesian solution to the inverse problem. Compared to most deterministic inverse techniques  Not just a single value for m, but a probability density which encapsulates all available information about the inverse problem.  Various summarizing statistics, e.g. mean, modes, marginal distribution and credible intervals, can be computed and quantify the uncertainty associated with a specific inverse solution.

Materials Process Design and Control Laboratory 8 Bayesian inference  Note that it is not necessary to compute the normalizing constant under most circumstances, e.g. using MCMC. Therefore, the posterior can be evaluated as follows:  Likelihood function : ― Example: deterministic forward problem G(m). Additive measurement error ― A simple model assumes that independent additive Gaussian random errors with mean zero and standard deviation σ account for the experimental noise. Then the likelihood is given by

Materials Process Design and Control Laboratory 9 Bayesian inference posterior density Likelihood function L (m) prior density Normalizing const  Prior : ― Incorporates additional information (e.g. physical constraints, smoothness, structure, expert judgment, etc). ― Can choose non informative prior.  Hierarchical Bayesian Models – hyper parameters θ, φ:

Materials Process Design and Control Laboratory 10 Example: Source Inversion Source described by location parameters m, active for. Strength s = 5.0, width τ = 0.1. Data are collected at 25 sensors on a regular 5x5 grid at two times t = 0.05 and t = 0.1. Measurement error is additive Gaussian. Priors: Goal : Given a set of noisy temperature T, try to estimate the unknown source location m

Materials Process Design and Control Laboratory 11 Posterior density  A hierarchical Bayesian model is augmented to detect the noise level σ.  A standard practice to select priors for hyper-parameters is to use conjugate priors. Thus, an inverse Gamma distribution is chosen for σ 2. where (α,β) is the parameter pair of the inverse Gamma distribution.  The posterior is non-standard and implicit, and the dimension of the posterior state space may be high. Thus, numerical sampling methods are needed for its exploration.  The most popular numerical method for this purpose is Markov Chain Monte Carlo.

Materials Process Design and Control Laboratory 12 Markov Chain Monte Carlo  In this problem, the posterior density is sampled using the following hybrid of the Metropolis-Hastings algorithm and the Gibbs sampler:  The acceptance ratio is defined as  The full conditional is also an inverse Gamma

Materials Process Design and Control Laboratory 13 Markov Chain Monte Carlo  Note that, in the traditional Bayesian formulation, if the underlying forward problem is a linear problem, then it is easy to derive the analytical form of the posterior density.  However, here, the forward problem is a non-linear problem due to the source term. Therefore, the evaluation of the acceptance ratio at each MCMC iteration requires computing the likelihood for the proposed move, which in turn involves solving the forward problem at each sample in a deterministic space.  Since each forward solver itself can be computationally expensive for complex models, the solution time can quickly become very large.  In the classical MCMC method, we do not take fully into account the stochastic nature of the forward model, i.e. that the system response is actually a stochastic function defined in the sample space. Next, we introduce a novel way to calculate the likelihood function through probability theory.

Materials Process Design and Control Laboratory 14 Propagate prior uncertainty  In Bayesian problem, everything is assumed random. Thus, the forward problem becomes a stochastic partial differential equation (SPDE). The input uncertainty is the assumed prior density.  Based on the idea of uncertainty quantification, we can propagate the input uncertainty through the forward model and construct a stochastic surrogate model for the forward problem to reduce the computational time  The basic idea is to construct an interpolant in the stochastic prior space using adaptive sparse grid collocation method. Then the interpolant can be used as a surrogate model for the stochastic forward model.

Materials Process Design and Control Laboratory 15 Mathematical preliminaries Ω Sample space of elementary events Real line Random variable  MAP Collection of all possible outcomes Each outcome is mapped to a corresponding real value Interpreting random variables as functions A general stochastic process is a random field with variations along space and time – A function with domain (Ω, Τ, S) Stochastic prior state space  Define and their images:  Then the state space of is defined as with the joint PDF denotes as.  For example, if is an independent uniform random variable in, then  Define a complete probability space with a sample space which corresponds to the outcomes of some experiments, being the σ-algebra of subsets in (these subsets are called events) and the probability measure.  In this framework: image:

Materials Process Design and Control Laboratory 16 Stochastic forward problem 1  Without loss of generality, the data d is the solution u to the forward problem at given sensor locations.  We can define a complete probability space, where the sample space is the set of all possible outcomes of. The forward problem is now taken as a stochastic model. Find a solution such that for P-almost everywhere, the following equation holds: for all  The random vector has a state space and a joint PDF which is the prior PDF. Therefore, for, the realization of random vector takes value in. We define as the stochastic prior state space.  The stochastic posterior state space can be also defined as where the event space is the inverse image (under ) of. 1. X. Ma, N. Zabaras, An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method, Inverse Problems 25 (2009) (27pp).An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method

Materials Process Design and Control Laboratory 17 Stochastic forward problem  Since the unknown parameters are always discretized into a random vector, then by using the Doob-Dynkin lemma, u is also a function of i.e.. Also, define D as a d-dimensional physical domain. Therefore, we can restate the deterministic forward problem as the following stochastic forward problem: find a function, such that the following holds:  For each realization, the function value gives one realization of the predicted data, which is equivalent to the solution of the deterministic forward problem using the same m as the input. In this way, the repetitive solution of the deterministic forward problem in MCMC is substituted with the solution to the stochastic forward problem.  When the prior space is the unbounded, we can always truncated it to a bounded one based on some prior information about the unknowns. The only requirement is that the prior space is large enough to contain the posterior space completely.

Materials Process Design and Control Laboratory 18 Sparse grid collocation method 1 1. X. Ma, N. Zabaras, A hierarchical adaptive sparse grid collocation method for the solution of stochastic differential equations, JCP, in press.in press  In the next few slides, a novel way based on sparse grid collocation method to find the stochastic surrogate model is reviewed.  The basic idea of this method is to have a finite element approximation for the spatial domain and approximate the multi-dimensional stochastic space using interpolating functions on a set of collocation points  Suppose that we can find a finite element approximate solution u to the deterministic solution of the problem for each realization ; we are then interested in constructing an interpolant of u by using linear combinations of the solutions.  The interpolation can be constructed by using either a full-tensor product of 1D interpolation rule of the so-called sparse grid interpolation method based on the Smolyak algorithm.

Materials Process Design and Control Laboratory 19 Collocation based framework Function value at any point is simply Stochastic function in 2 dimensions Need to represent this function Sample the function at a finite set of points Use basis functions to get a approximate representation Spatial domain is approximated using a FE, FD, or FV discretization. Stochastic domain is approximated using multidimensional interpolating functions

Materials Process Design and Control Laboratory 20 From one-dimension to higher dimension  Denote the one dimensional interpolation formula as with the set of support nodes:  In higher dimension, a simple case is the tensor product formula For instance, if M=10 dimensions and we use k points in each direction Number of points in each direction, k Total number of sampling points x x x10 10 This quickly becomes impossible to use. One idea is only to pick the most important points from the tensor product grid.

Materials Process Design and Control Laboratory 21 Choice of collocation points and nodal basis  In the context of incorporating adaptivity, we have utilized the Newton- Cotes grid using equidistant support nodes. The number of nodes is defined as, if ; else Then the support nodes are  Furthermore, by using the linear hat function as the univariate nodal basis function, one ensures a local support in contrast to the global support of Lagrange polynomials. This ensures that discontinuities in the stochastic space can be resolved. The piecewise linear basis functions can be defined as, for, and for and

Materials Process Design and Control Laboratory 22 From nodal basis to hierarchical basis  Using the nested property of the grid points, we can rewrite the 1D formula in a hierarchical fashion. We define. With and, we obtain sincewe have  For simplifying the notation, we consecutively number the elements in, and denote the j-th points of as hierarchical surplus

Materials Process Design and Control Laboratory 23 Nodal basis vs. hierarchical basis Nodal basis Hierarchical basis

Materials Process Design and Control Laboratory 24 Multi-dimensional hierarchical interpolation  The N-dimensional multilinear basis functions can be defined as  Furthermore, through a new multi-index set we can define the hierarchical basis as  Using the 1D formula, the sparse interpolant, where is the depth of sparse grid interpolation and is the number of stochastic dimensions, is given by the Smolyak algorithm as Here we define the hierarchical surplus as:

Materials Process Design and Control Laboratory 25 Adaptive sparse grid collocation (ASGC) Let us first revisit the 1D hierarchical interpolation  For smooth functions, the hierarchical surpluses tend to zero as the interpolation level increases.  On the other hand, for non-smooth function, steep gradients/finite discontinuities are indicated by the magnitude of the hierarchical surplus.  The bigger the magnitude is, the stronger the underlying discontinuity is.  Therefore, the hierarchical surplus is a natural candidate for error control and implementation of adaptivity. If the hierarchical surplus is larger than a pre-defined value (threshold), we simply add the 2N neighbor points of the current point.

Materials Process Design and Control Laboratory 26 Adaptive sparse grid collocation: Algorithms  Let be the parameter for the adaptive refinement threshold.

Materials Process Design and Control Laboratory 27 Adaptive sparse grid interpolation Ability to detect and reconstruct steep gradients

Materials Process Design and Control Laboratory 28 ASGC in Bayesian inference  By using the ASGC method, the stochastic solution to the stochastic forward problem can now be approximated by the following reduced form:  To obtain the predicted data for any, instead of solving the deterministic forward problem, we can simply substitute into above equation to compute the value of within a certain accuracy. Furthermore, we can easily compute the value of the likelihood.  In this case, most of the computational time is spent on constructing the interpolant for the solution. This construction is embarrassingly parallel and thus the computational time is minimal. Once the interpolant is constructed, we can simply store it for future use when new observed data are available.  We only consider the unit hypercube here. Other shapes of the bounded prior space can be easily transformed to the unit hypercube.

Materials Process Design and Control Laboratory 29 Solution to the stochastic forward problem  Return to the 2D source inversion problem. The stochastic forward problem is solved in the prior space by ASGC.  Propagate the uniform prior on source location. First, we need to check the convergence and accuracy of the surrogate model. Surface response of a single component of the stochastic forward problem solution at point (0.5,0.5) and t=0.05 with different ε.

Materials Process Design and Control Laboratory 30 Solution to the stochastic forward problem  The accuracy of the surrogate model can be interrogated in several ways. First we compare the PDF at t = 0.05 and t = 0.1 at point (0.5,0.5) with the direct computation by solving the deterministic forward problem.  Therefore, it is seen that when, the PDF converges to the solution by Direct computation, which verifies the accuracy of the surrogate stochastic forward model.

Materials Process Design and Control Laboratory 31 Solution to the stochastic forward problem  Next, we examine the accuracy of the surrogate likelihood defined as  The exact source location is (0.5, 0.5) with 5% noisy in the data. Contours of the exact (solid) and the surrogate likelihood (dashed) Convergence of the surrogate posterior: Kullback-Leibler distance vs. ε

Materials Process Design and Control Laboratory 32 Solution to the inverse problem  Now we illustrate the usage of the surrogate mode for the Bayesian inference approach on the model inverse problem.  The initial guess for m and σ is (0,0) and 1.0. The pair of parameters (α,β) for the inverse Gamma distribution is taken to be (1x10 -3, 1x10 -3 ).  The proposal distribution in the Metropolis-Hastings algorithm is a random walker sampler  Unless otherwise specified, the scale parameter is taken to be, and the surrogate by stochastic collocation method is constructed with  The burn-in period of the Markov chain is 10000, and the last realizations are used to compute relevant statistical quantities.

Materials Process Design and Control Laboratory 33 Trace plots and autocorrelation plot of the Markov Chain  The acceptance rate is  Visual inspection suggests the chain mixes very well.  The current choice of scale parameter in the proposal distribution leads to a fast decay with lag along the chain, concurring with good mixing in the trace plot.

Materials Process Design and Control Laboratory 34 Speed up of the surrogate model  The motivation of using this method is to speed up the Bayesian inference for nonlinear problem, where repeated solution of the forward model for each MCMC sample is replaced by evaluation of. Time : 34 hours The direct FEM computation for the likelihood needs evaluations Time : 84.18s to construct the surrogate model on 20 nodes and 26.9s for a single MCMC chain The number of collocation points (thus the number of direct computation needed) in the sparse grid is 1081.

Materials Process Design and Control Laboratory 35 Marginal density  A distinct feature of MCMC is easy extraction of the marginal distribution for the component of interest via kernel density estimation (KDE).  Although the prior distribution is only uniform, the posterior density of greatly refines the prior distribution.  It is seen that the range of distribution is much larger with 5% noise, whereas it is much more concentrated around the exact value with 1% noise.  From the tables, it can be seen that the current Bayesian hierarchical formulation can successfully detect the noise levels.

Materials Process Design and Control Laboratory 36 Effect of source locations  Besides the great computational savings, another advantage of using this method is that it is reusable when there are new measurements coming after constructing the surrogate model.  The following table shows the results for four other source locations: (0.1,0.9), (0.75,0.25) and two other random draws from the prior with 1% noise in the data. The same surrogate model is used.  It can be seen that no matter where the source location is, the method can always infer the exact value without performing any additional direct FEM computations as in the traditional MCMC method.

Materials Process Design and Control Laboratory 37 Error with different ε and noise level  It is also interesting to compare the results with respect to different ε and noise level. The true source location is chosen at (0.2,0.8). The error is defined as the maximum error between the posterior mean and the true source locations.  Surprisingly, for a given noise level, even a big threshold can give us rather accurate results. This is possibly due to error cancellations between the error of the surrogate model and the measurement error.

Materials Process Design and Control Laboratory 38 Effect of the prior distribution  We next demonstrate that the accuracy of the method is not affected by the prior distribution provided the prior space is large enough to contain the posterior space.  The prior is chosen as independent Gaussian, which is a special of MRF. The hierarchical PPDF is  It is obvious that the unit square contains the posterior space. So, we truncate the unbounded space to the unit square and we do not need to perform the calculation using ASGC again.

Materials Process Design and Control Laboratory 39 Effect of the size of prior space  We conduct several computations with different sizes of prior space. The surrogate model is constructed with and 5% noise in the data. The true source location is chosen as (0.2, 0.8).  As long as the prior space contains the posterior space, we can always obtain accurate posterior mean estimation. Increasing does not affect the accuracy but it affects the computational cost. The number of collocation points increases with larger and also the burn-in length of the MCMC chain is elongated. Therefore, it is important to choose an appropriate space which balances accuracy and computational cost.

Materials Process Design and Control Laboratory 40 Example : Permeability estimation  In this example, we illustrate the method on the nonlinear inverse problem of estimating the spatial permeability field. The pressure and velocity and characterized by the following set of dimensionless equations: where denotes the source/sink term. injection well production well  To impose the non-negativity of the permeability, from now on we will treat the logarithm of the permeability as the main unknown of our problem.  To generate simulation data, we consider smooth permeability of the following form Pressure is measured at a 5x5 sensor network and 5% noise is added.

Materials Process Design and Control Laboratory 41 MRF and process convolution model  In the general Bayesian setting, the unknown permeability field and the forward model must be discretized. If the FE method is used, the field is discretized onto the nodes of the FE mesh. Denote by the number of nodes of the mesh; then we can write both the prior and posterior densities in terms of the unknown parameter  Possible choices for the prior of is the Gaussian process (GP) and the Markov Random Field. A Gaussian MRF is of the form Number of neighbors of site if and are adjacent.  However, it is noted that the dimensionality of is generally very high since is above 1000 in a typical finite element mesh. Exploring the posterior state space using MCMC is not trivial and the acceptance rate is generally very low. To this end, we use the process convolution approach to reduce the dimensionality.

Materials Process Design and Control Laboratory 42 MRF and process convolution model  The GP at any location can be represented as The kernel function is chosen to be a mean zero Gaussian kernel at a discrete set of points which controls the spatial structure of the underlying process and are independent  Thus the prior density of is  The inverse problem has now been transformed to an inference problem on the coefficients  If the locations of the kernels are at a regular grid, a MRF can also be used for the prior distribution of the underlying process on this grid.  The standard deviation of the kernel is usually equal to the distance between adjacent locations of the kernels and the lattice grid of the kernel locations is usually larger than the space of the observed data in order to eliminate inaccurate results near the boundaries.

Materials Process Design and Control Laboratory 43 Permeability estimation: fixed kernel locations  If the unknown is also denoted as, then the hierarchical PPDF can be computed as follows:  We choose a 5 x 5 lattice in as kernel locations to construct the discrete process convolution model which corresponds to a total of 25 kernels, where the Gaussian kernel is used. MRF is used as the prior and the stochastic prior space is chosen.  A 25-dimensional ASGC with is used, which has collocation points. However, the corresponding number of collocation points for a conventional sparse grid is Thus, the advantage of using ASGC is obvious.

Materials Process Design and Control Laboratory 44 Permeability estimation: fixed kernel locations 5% noise in the data TruePosterior mean Marginal densities

Materials Process Design and Control Laboratory 45 Permeability estimation: fixed kernel locations 1% noise in the data 5 x 5 sensor grid 9 x 9 sensor grid Marginal of each kernel strengths 5% noise 1% noise

Materials Process Design and Control Laboratory 46 Random kernel locations  are independent Gaussian random variables. Each kernel has three degrees of freedom, i.e. and x, y locations. Posterior mean: 4 kernelsPosterior mean: 5 kernels

Materials Process Design and Control Laboratory 47 Conclusions  An efficient Bayesian approach is introduced based on adaptive sparse grid collocation method. Efficient propagation of the prior uncertainty through the forward model Inference of spatial fields, with dimensionality reduction using MRF process convolution model The accuracy of the methods is assessed on two non-linear inverse problems. The speed up of this method comparing with that of direct computation is drastically.  However, there are also some limitations of the method. ASGC relies exclusively on a bounded input support. Therefore, it is crucial to choose appropriate and. As long as the prior space includes the support of the posterior, the method provides very good results.

Materials Process Design and Control Laboratory 48 References 1. X. Ma, N. Zabaras, An efficient Bayesian inference approach to inverse problems based on an adaptive sparse grid collocation method, Inverse Problems 25 (2009) (27pp).25 (2009) (27pp). 2. X. Ma, N. Zabaras, A hierarchical adaptive sparse grid collocation method for the solution of stochastic differential equations, Journal of Computational physics, in press, 2009.in press,